Linguistics with CLARIN
OpenSONAR
Jan Odijk
LOT Winterschool
Amsterdam, 2015-01-13
1
Overview
•
•
•
•
SONAR
OpenSONAR
Methodological Considerations
Google?
2
Overview
 SONAR
• OpenSONAR
• Methodological Considerations
• Google?
3
SONAR
• SONAR Dutch corpus
• 500 million tokens
• Written language
• (for spoken language: CGN)**
• Many different text types
• Includes `new media‘ (sms, tweets, blogs,
…)
• but not balanced (mainly because of legal
restrictions)
4
SONAR
•
•
•
•
•
•
•
FoLIA Format
Pos, lemma, word properties for each token
`metadata’ for each document
Created in the STEVIN-project (2004-2011)
Can be obtained via TST-Centrale
Reference:
Oostdijk Nelleke, Martin Reynaert, Véronique Hoste,
Ineke Schuurman (2013). `The Construction of a 500Million-Word Reference Corpus of Contemporary
Written Dutch’, In [Spyns & Odijk 2013]. [pdf]
5
SONAR
• Some interesting Annotated Text Corpora
• English
• British National Corpus
• Corpus of Contemporary American English (and many more
at BYU)
• American National corpus
• Multiple languages
• CHILDES Corpora
• German
• Das Deutsche Referenzkorpus
6
SONAR
• Some interesting Annotated Text Corpora
• Spanish
• Syntactic Spanish Database (SDB) University of Santiago de Compostela.
160,000 clauses / 1.5 million words.
• Ancora-ES (and Ancora-CA) and others
• Panacea Annotated Corpus (downloadable)
• Corpus Molinero (but no annotations)
• Corpus Tecnic de l’IULA
• Dutch
•
•
•
•
Corpus Gesproken Nederlands (CGN)
SONAR en SONAR Nieuwe Media
VU-DNC
Discan
• …
7
Overview
• SONAR
 OpenSONAR
• Methodological Considerations
• Google?
8
OpenSONAR
• Search interface to the SONAR Corpus
• Some Interfaces to Corpora for other lgs:
Interface
BNCweb interface at Lancaster
IMS Open Corpus Work Bench
Corpus of Contemporary American
English
Corpus of Contemporary Dutch
TrovA
Språkbanken
Corpuscle
Bwananet
Language(s)
British English
German
American English
Dutch
Multiple
Swedish
Norwegian
Spanish, Catalan, ..
9
OpenSONAR
• Search interface to the SONAR Corpus
• Runs on INL (Instituut voor Nederlandse Lexicologie),
one of the Dutch CLARIN Centres
• http://opensonar.clarin.inl.nl/
• Login with the account of your institute
• Federated login, single sign on (CLARIN)
• created in the CLARIN-NL project (2009-2014)
• Available since November 2014 (!)
10
OpenSONAR
• Back-end based on BlackLab, developed at INL
– Open Source Software, based on Apache Lucene
– https://github.com/INL/BlackLab#readme
– https://github.com/INL/BlackLab/wiki/BlackLab-blog
• Front-end developed by UvT, `Whitelab’
– Open Source Software
– https://github.com/INL/WhiteLab
11
OpenSONAR
• 4 interfaces
• Simple, extended, advanced, expert
• Expert = CQP language (CQL)
• Grouping, Restricting by metadata
• Pos-codes:
• Van Eynde, Frank (2004), `Part Of Speech
Tagging en Lemmatisering Van Het Corpus
Gesproken Nederlands’, Centrum voor
Computerlinguïstiek, K.U.Leuven [pdf]
12
OpenSONAR
• See Scenario demo OpenSONAR
13
Overview
• SONAR
• OpenSONAR
 Methodological Considerations
• Google?
14
Methodological
Considerations
• Performance (actually used) data
• Including errors, hesitations, fillers, etc
• Good for certain research questions
• Less good for other research questions
• No `negative’ data
– Linguists sometimes want to know what is NOT
possible in language
– More difficult to find non-standard examples (e.g.
examples not covered by the grammar used for a
treebank)
15
Methodological
Considerations
• Danger of circularity
• ‘Which verbs occur with a predicative adjective?’
• the verbs that have been specified as such in the
grammar underlying a treebank
• Can be avoided by globally knowing how the relevant
grammar works
• No controlled experiments
– Minimal pairs seldom occur naturally
– BUT: Corpora/Treebanks can be used to construct
minimal pairs on the basis of really occurring
examples
16
Methodological
Considerations
• Annotations have mainly been made by
automatic programs
• They make errors
• `absurd errors’
• Insufficient information errors
• People also make errors but different
ones
• `sloppiness errors’
17
Methodological
Considerations
• Large corpora:
• high frequency results are more reliable
results
• low frequencies are suspect
• Small corpora:
• human verification and correction is
required
18
Methodological
Considerations
• Desired:
• get all relevant examples (high recall)
• no or few irrelevant examples (high
precision)
• Very difficult to achieve
• Critical analysis of the results is always
required
19
Methodological
Considerations
• User friendly interface implies limitations:
– Cf. OpenSONAR interface (advanced: no extended
pos (inflectional information)
– Several examples can be given for GrETEL
20
Methodological
Considerations
• Simple cases can be solved by small
adaptations in the query, e.g.
• Start with the graphical interface
• Adapt in the expert interface
• Adapting easier than creation from
scratch
21
Overview
• SONAR
• OpenSONAR
• Methodological Considerations
 Google?
22
Google?
Property
Google
What you want
String search
yes
yes
Relation between strings
nearness
Grammatical relations
Search for function words
No / unreliable
Yes
Search for morphosyntactic and syntactic
properties
no
Yes
23
Google?
Property
Google
What you want
Search within a sentence,
paragraph?
No (documents only)
Sentence, paragraph,
section etc
results
List of documents
List of sentences,
paragraphs, sections,
documents
Grouped /sorted
(analyzed) results
no
yes
Construction search
no
Yes
Single language only
unreliable
Yes
Size
huge
Huge (but so far there is
only small (1m tokens) to
large (500m tokens)
24
Thanks for your Attention!
25